Top-down feedback does not benefit speech recognition; on the contrary, it can hinder it. No experimental data imply that feedback loops are required for speech recognition. Feedback is accordingly unnecessary and spoken word recognition is modular. To defend this thesis, we analyse lexical involvement in phonemic decision making. TRACE (McClelland & Elman 1986), a model with feedback from the lexicon to prelexical processes, is unable to account for all the available data on phonemic decision making. The modular Race model (Cutler & Norris 1979) is likewise challenged by some recent results, however. We therefore present a new modular model of phonemic decision making, the Merge model. In Merge, information flows from prelexical processes to the lexicon without feedback. Because phonemic decisions are based on the merging of prelexical and lexical information, Merge correctly predicts lexical involvement in phonemic decisions in both words and nonwords. Computer simulations show how Merge is able to account for the data through a process of competition between lexical hypotheses. We discuss the issue of feedback in other areas of language processing and conclude that modular models are particularly well suited to the problems and constraints of speech recognition.

Norris, McQueen & Cutler claim that all known speech recognition data can be accounted for with their autonomous model, “Merge.” But this claim is doubly misleading. (1) Although speech recognition is autonomous in their view, the Merge model is not. (2) The body of data which the Merge model accounts for, is not, in their view, speech recognition data.

I present two criticisms: (1) Merge is a model of performance in speech perception experiments but not an ecologically valid model integrating both word recognition and speech perception, and (2) Merge's implementation of the variability of lexical effects must be made more precise or the model is indistinguishable from its alternatives.

The present description of the Merge model addresses only auditory, not audiovisual, speech perception. However, recent findings in the audiovisual domain are relevant to the model. We outline a test that we are conducting of the adequacy of Merge, modified to accept visual information about articulation.

We consider the motivation for the principle of bottom-up priority and its consequence for information flow in Merge. The relationship between the bottom-up priority rule and inhibitory effects is also discussed, along with data that demonstrate inhibitory influences in phoneme monitoring.

Our commentary outlines a number of arguments questioning an autonomous model of word recognition without feedback. Arguments are presented against the need for a phonemic decision stage and in support of a featural level in a model including feedback.

I respond to Norris et al.'s criticism of Gaskell and Marslen- Wilson (1997). When the latter's network is tested in circumstances comparable to the Merge simulations in the target article, it produces the desired pattern of results. In another area of potential feedback in spoken word processing, aspects of lexical content influence word recognition and our network provides a simple explanation of why such effects emerge. It is unclear how such effects would be accommodated by Merge.

The Merge model suggests that lexical effects in phonemic processing reflect the activation of post-lexical phonemic representations that are distinct from prelexical phonemic input representations. This distinction seems to be unmotivated; the phoneme fails to capture the richness of prelexical representation. Increasing the information content of input representations minimizes the potential necessity for top-down processes.

The model of phoneme monitoring proposed by Norris et al. is implausible when implemented in a localist connectionist network. Lexical representations mysteriously inform phoneme decision nodes as to the presence or absence of a target phoneme.

The brain contains ubiquitous reciprocal bottom-up and top-down intercortical and thalamocortical pathways. These resonating feedback pathways may be essential for stable learning of speech and language codes and for context-sensitive selection and completion of noisy speech sounds and word groupings. Context-sensitive speech data, notably interword backward effects in time, have been quantitatively modeled using these concepts but not with purely feedforward models.

Norris, McQueen & Cutler demonstrated that feedback is never necessary during lexical access and proposed a new autonomous model, that is, the Merge model, taking into account the known behavioral data on word recognition. For sentence processing, recent event-related brain potentials (ERPs) data suggest that interactions can occur but only after an initial autonomous stage of processing. Thus at this level too, there is no evidence in favor of feedback.

Recent investigations indicate that, around 7-months-of-age, infants begin to show some ability to recognize words in fluent speech. In segmenting and recognizing words, infants rely on information available in the speech signal. We consider what implications these findings have for adult word recognition models in general, and for Merge, in particular.

The massive acoustic redundancy of minimally contrasting speech sounds, coupled with the auditory integration of psychoacoustically similar acoustic properties produces a highly invariant percept, which cannot be improved by top-down feedback from the lexicon. Contextual effects are also bottom-up but not all entirely auditory and may thus differ in whether they affect sensitivity or only response bias.

We applaud Norris et al.'s critical review of the literature on lexical effects in phoneme decision making, and we sympathize with their attempt to reconcile autonomous models of word recognition with current research. However, we suggest that adaptive resonance theory (ART) may provide a coherent account of the data while preserving limited inhibitory feedback among certain lexical and sublexical representations.

Norris et al. argue against using evidence from phonetic decision making to support top-down feedback in lexical access on the grounds that phonetic decision relies on processes outside the normal access sequence. This leaves open the possibility that bottom-up connectionist models, with some contextual constraints built into the access process, are still preferred models of spoken-word recognition.

Our long-standing hypothesis has been that feedforward information flow is sufficient for speech perception, reading, and sentence (syntactic and semantic) processing more generally. We are encouraged by the target article's argument for the same hypothesis, but caution that more precise quantitative predictions will be necessary to advance the field.

A comparison of Merge, a model of comprehension, and WEAVER, a model of production, raises five issues: (1) merging models of comprehension and production necessarily creates feedback; (2) neither model is a comprehensive account of word processing; (3) the models are incomplete in different ways; (4) the models differ in their handling of competition; (5) as opposed to WEAVER, Merge is a model of metalinguistic behavior.

Norris, McQueen & Cutler argue that there is no need for feedback in word recognition. Given the accumulating evidence in favor of feedback as a general mechanism in the brain, I will question the utility of a model that is at odds with such a general principle.

Approaches to model evaluation in terms of Occam's razor or principles of parsimony cannot avoid judgements about the relative importance of aspects of the models. Assumptions about “core processing” are usually considered more important than those related to decision components, but when the decision is related to a central feature of the processing, it becomes extremely difficult to tease apart the influences of core and decision components and to draw sensible conclusions about underlying architecture. It is preferable, where possible, to use experimental procedures that avoid the necessity for subject decisions related to critical aspects of the underlying process.

As a proponent of phoneme-like units in speech perception, I am very sympathetic to Merge's use of phoneme-oriented input. However, in the absence of any known way to provide input in exactly the form assumed, further consideration needs to be given to how the variation in the details of the specification of this input might affect Merge's (and Shortlist's) overall behavior.